AI in Education: Beyond Tool Application

• AI in Education is a multidimensional paradigm integrating automation with human agency to facilitate reflective and hybrid learning models.
• The critique of the tool-centric approach exposes risks like dehumanization and reduced learner agency, urging a shift towards co-adaptive educational systems.
• Formal frameworks—externalization, internalization, and extension—offer practical methods to balance automation with human insight in transforming educational processes.

AI in education has traditionally been conceptualized within a narrow paradigm of tool application—automating discrete instructional tasks or augmenting classroom resources. Current research demonstrates that such a reductionist framing is insufficient to capture the depth and breadth of AI's potential to transform educational systems, processes, and epistemic cultures. The most robust contemporary frameworks instead articulate a multidimensional vision of AI as integral to co-constructing, mediating, and amplifying human cognitive activity, not merely supplanting it with algorithmic automation (Cukurova, 2024).

1. Critique of the Tool-Centric Paradigm

The "AI as tool" model centers on the replacement or automation of specific cognitive or pedagogical tasks: decision-making, content delivery, feedback, or assessment. In this approach, AI systems are data-driven prediction engines (e.g., LLMs as "stochastic parrots"), characterized by mappings $X \rightarrow Y$—student data in, pedagogical output out, with human agency relegated to data provision or consumption of system outputs. Over-reliance on this mode risks several adverse outcomes:

• Dehumanization of Learning: Socio-cultural, emotional, and metacognitive dimensions are marginalized, resulting in rigid instructional workflows that neglect tacit knowledge, creativity, and the transformative potential of lived educational experience.
• Agency Erosion: Learners and educators increasingly occupy passive, reactive positions, simply responding to automated feedback or system decisions, rather than actively negotiating learning trajectories.
• Neglect of Hybrid and Analogy-Oriented Roles: The formative tradition in AI in Education (AIED) that leveraged computation as a lens to understand human learning—and vice versa—has been overshadowed by tool-centric operationalism, undermining an analytical relationship between educational theory and AI system design (Cukurova, 2024).

2. Formal Framework: Three Complementary Conceptualizations

To move beyond automation, AI in education is best understood along two axes: degree of automation and degree of human agency/control. This yields three principal modes, each defined with explicit formal notation:

a. Externalization ("AI as Tool")

• Definition: AI automates or replaces a subset of human cognitive tasks ($T_H \subseteq H$).
• Formalization: AI system $A_{\text{external}}: X \rightarrow Y$, where $H_{\text{system}} = H \setminus T_H$.
• Properties: High automation, low human agency.

b. Internalization ("AI as Mental Model")

• Definition: AI-based outputs (e.g., dashboards, computational models) serve as triggers for human reflection, updating internal mental models.
• Formalization: For current mental model $mm_n$, given an AI construct $M: X\rightarrow Z$ and human-interpretable mapping $\varphi$, learning is:

$$mm_{n+1} = mm_n + \gamma\,(\varphi(M(x)) - mm_n), \quad 0 < \gamma \leq 1$$

• Properties: Low automation, high human agency. Emphasizes epistemic agency, metacognition, and human interpretation.

c. Extension ("Hybrid Intelligence")

• Definition: Human and AI form a bidirectional, tightly-coupled hybrid system with emergent properties.
• Formalization: Synergy $S$ is quantified as

$S = I(H \otimes A) - [I(H) + I(A)] \geq 0$

where $I(\cdot)$ measures intelligence or task performance, and $\otimes$ denotes a coupling operator such as real-time interactive dialogue.

• Properties: High automation, high human agency. The objective is $S \gg 0$—superadditive synergy in learning performance, adaptability, or problem-solving.

3. Illustrative Applications and Distinguishing Features

Each conceptualization entails distinct system architectures, interaction patterns, and pedagogical affordances:

Conceptualization	Example Systems	Learner Interaction	Design/Pedagogy
Externalization	ITS (Cognitive Tutor, Duolingo, ASSISTments),	Largely reactive, data provision	Rule/data-driven automation,
	Multimodal engagement detectors, automated lectures		minimal input beyond data
Internalization	Collaboration dashboards, engagement visualizations	Exploratory, human-driven action	Interpretability, transparency, reflection
Extension (Hybrid)	Adaptive dialogical agents, human-AI co-design	Bidirectional negotiation, co-action	Co-adaptive loops, fading AI support

• Externalization: Scalability and consistency in structured domains like mathematics and language. Risk: decontextualized, dehumanized education and atrophy of core human skills.
• Internalization: Enhanced metacognition and conceptual change. Risk: technical challenge of model validity, risks of externally defined "normativity."
• Extension: Potential for emergent, distributed intelligence. Risk: implementation lag, possible accidental dominance of automation if coupling is not balanced.

4. Empirical and Theoretical Benefits and Risks

A systematic analysis exposes the affordances and vulnerabilities of each approach (Cukurova, 2024):

Conceptualization	Benefits	Limitations / Risks
Externalization	Scalability, proven gains in well-defined domains	Dehumanizes learning, low human agency, atrophy
Internalization	Metacognition, high learner ownership, flexibility	Model accuracy, normativity of "good" outputs
Extension (Hybrid)	Superadditive performance, preserved agency	Research gap, risk of over-automation

There is particular peril in over-emphasizing externalization: critical competencies may atrophy when learning is reduced to consumption of AI-driven outputs.

5. Design Principles and Recommendations for Practice

Transcending tool application requires an intentional, context-sensitive blending of the three conceptualizations:

• Multiplexed Deployment: Align AI's role—externalizer, model for internalization, or hybrid partner—with explicit pedagogical goals. Do not default to maximized automation.
• AI Literacy Curriculum: Develop frameworks (e.g., aligned to UNESCO AI Competency models) for both teachers and students, ensuring clear understanding of AI's functional limits, epistemic biases, and interpretive demands.
• Socio-Technical Alignment: Center design processes on active co-design with stakeholders, embedding “human-in-the-loop” override and interpretability at critical junctures.
• Assessment Innovation: Move beyond final-product evaluation to valorize reflective processes, self-regulation, and spaced engagement analytics.
• Hybrid Prototyping: Advance research on human–AI collaboration (e.g., adaptive dialogical agents), implementing scaffolding that fades as human competence increases (maximizing learning rate $\gamma$ only to the degree that it catalyzes beneficial internalization).
• Longitudinal Monitoring: Systematically evaluate hybrid configurations for both augmentation and skill atrophy, embedding longitudinal study designs into system rollouts.

6. Broader Theoretical and Systemic Implications

The expanded vision for AI in education integrates fundamental insights from cognitive science, epistemology, and socio-cultural learning theory:

• Certain core elements of deep learning—reflection, lived experience, and the emergence of tacit or situated knowledge—resist full algorithmic formalization or automation.
• The future of AIED must entail not only system design, but educational innovation that prepares learners to operate ethically, reflexively, and critically in a world saturated by AI.
• The most ambitious agenda asks: "When do we amplify human intelligence? When do we risk atrophying it? And how can AI models serve as objects for thinking about our own intelligence?"

Education systems must thus move toward hybrid wisdom—where automation, interpretability, and co-adaptive hybrid intelligence are not mutually exclusive, but engineered in tandem to realize and sustain the deepest forms of human learning, agency, and flourishing (Cukurova, 2024).